1: \begin{abstract}

2: \begin{sloppy}

3: We discuss the computational complexity of random 2D Ising spin

4: glasses, which represent an interesting class of constraint

5: satisfaction problems for black box optimization. Two extremal

6: cases are considered: (1) the $\pm J$ spin glass, and (2) the

7: Gaussian spin glass. We also study a smooth transition between

8: these two extremal cases. The computational complexity of all

9: studied spin glass systems is found to be dominated by rare events

10: of extremely hard spin glass samples. We show that complexity of

11: all studied spin glass systems is closely related to Fr\'echet

12: extremal value distribution. In a hybrid algorithm that combines

13: the hierarchical Bayesian optimization algorithm (hBOA) with a

14: deterministic bit-flip hill climber, the number of steps performed

15: by both the global searcher (hBOA) and the local searcher follow

16: Fr\'echet distributions. Nonetheless, unlike in methods

17: based purely on local search, the parameters of these

18: distributions confirm good scalability of hBOA with local search.

19: We further argue that standard performance measures for

20: optimization algorithms---such as the average number of

21: evaluations until convergence---can be misleading. Finally, our

22: results indicate that for highly multimodal constraint

23: satisfaction problems, such as Ising spin glasses,

24: recombination-based search can provide qualitatively better

25: results than mutation-based search.

26: \end{sloppy}

27: \end{abstract}

28: